def learn_non_conf(df, global_models):
"""
Learn a model for all non-conference games
Could this be rolled into the learn_conf function easily?
"""
# Filter by conference - should this happen inside or outside the for loop?
dfc = df.loc[df.home_conf != df.away_conf].copy().dropna(axis=0)
dfc = dfc.drop(['home_conf', 'away_conf'], axis=1)
bayes_non_conf = dict()
model_cols = [
'sagarin_home', 'sagarin_away', 'kenpom_off_home', 'kenpom_def_home',
'kenpom_off_away', 'kenpom_def_away', 'moore_home', 'moore_away',
'score_diff'
]
for day in pd.date_range(df.game_date.min(), df.game_date.max()):
# start at min+1? I think that would avoid the cold-start issues encountered on days with no games
day = str(day.date())
df_conf = dfc.loc[(dfc.game_date == day),
model_cols] # rename linreg_cols to model_cols?
# other thought: have prev day outside of if, put isinstance statements in initial if?
if (len(df_conf) == 0):
try:
prev_day = str(
(pd.to_datetime(day) - pd.Timedelta(days=1)).date())
if isinstance(bayes_non_conf[prev_day], LinearRegression):
# use global lr?
bayes_non_conf[day] = global_models[day]
elif isinstance(bayes_non_conf[prev_day], BayesianRidge):
bayes_non_conf[day] = bayes_non_conf[prev_day]
except:
bayes_non_conf[day] = global_models[
day] # should only happen on first day
else:
bayes = BayesianRidge(normalize=True)
bayes.coef_ = global_models[day].coef_
X = df_conf.drop('score_diff', axis=1)
y = df_conf.score_diff
bayes_mod = bayes.fit(X, y)
bayes_non_conf[day] = bayes_mod
return bayes_non_conf
